Erlang-A

From WFM Labs

Erlang-A is the extension of Erlang-C that explicitly models caller abandonment. It replaces Erlang-C's infinite-patience assumption with an exponential distribution of caller patience and produces accurate staffing recommendations when abandonment is non-trivial. Erlang-A matters because Erlang-C systematically over-states required staffing in any operation where callers leave the queue before being answered — and most contact-center queues do, especially during peak intervals or service incidents.

The math originated with Conny Palm in 1946 and is sometimes called the Erlang-A or Palm/Erlang-A model. The formula has not been a standard contact-center planning tool until relatively recently because the closed-form is more involved than Erlang-C, but every modern WFM platform that includes Erlang-A as a model option benefits from materially better calibration in any queue with > 3% abandonment.

Definition

Erlang-A models the M/M/c+M queue: Markovian arrivals, Markovian service, c agents, exponential caller patience. Each arriving caller has an independent random patience time drawn from an exponential distribution with mean 1/θ — if their wait reaches the patience time, they abandon. θ is the patience rate; 1/θ is the mean caller patience.

Inputs: arrival rate λ, mean handle time AHT, agent count n, mean caller patience 1/θ.

Outputs: probability of waiting, probability of abandonment, expected wait conditioned on getting served, abandonment rate, and Service Level at any threshold.

Formula / mathematics

The Erlang-A formula is more involved than Erlang-C; the standard form expresses the steady-state queue distribution and derives metrics from it. Let a = λ × AHT (offered load), and define μ = 1/AHT, θ = 1/(mean caller patience).

The probability of abandonment for an Erlang-A queue is:

P(Ab) = (θ / λ) × (Expected number waiting) − a residual term

In practical use, Erlang-A is solved numerically: given inputs, the queue's steady-state probability vector is computed, and SL, abandonment, and ASA are read off.

The intuition for the result vs Erlang-C:

  • Required staffing. Erlang-A returns lower required-FTE than Erlang-C at the same SL target, because abandonment relieves the queue without requiring a new agent. Difference can be 5–20% depending on patience and target SL.
  • Predicted SL at a given staffing. Erlang-A returns higher SL than Erlang-C, for the same reason — abandons make the queue effectively smaller.
  • Abandonment rate. Erlang-C predicts zero. Erlang-A returns a real number.

The relationship to caller patience matters: at very long mean patience (1/θ → ∞), Erlang-A converges to Erlang-C. At very short patience (impatient callers), Erlang-A predicts heavy abandonment and very different staffing.

Practitioner use

Erlang-A is appropriate when:

  • Expected abandonment exceeds ~3% of offered contacts.
  • The operation tracks abandonment as an explicit business metric (most modern operations do).
  • Caller patience can be estimated empirically from ACD data — the median time-to-abandon in past intervals is a reasonable proxy.

Estimating mean caller patience:

  • The right number is the mean of the patience distribution, not the mean actual wait or the mean wait of abandoned callers.
  • It can be estimated from observed abandonment behavior using Kaplan-Meier-style survival curves on abandon-time data, or — for first-pass planning — set it equal to the median time-to-abandon for the queue from recent intervals.
  • Typical values: 30–120 seconds for routine inbound voice; longer (90–300 s) for support or technical queues where callers expect to wait; much shorter (10–30 s) for high-volume / low-engagement queues.

The standard staffing workflow with Erlang-A:

  1. Forecast volume and AHT.
  2. Estimate mean caller patience from recent ACD data.
  3. Solve for minimum n such that SL(T) ≥ target and abandonment ≤ acceptable.
  4. Apply Shrinkage gross-up.

The dual-target structure (SL and abandonment) is why Erlang-A is more honest than Erlang-C: it makes the trade-off explicit. Erlang-C presents a single SL number and hides the abandonment cost; Erlang-A surfaces both.

When to use Erlang-A vs Erlang-C

  • Use Erlang-C when expected abandonment is < 3%, the operation has stable patience, and a single-target SL is sufficient.
  • Use Erlang-A when expected abandonment is non-trivial, the operation has reliable patience data, and abandonment rate is itself a business metric.
  • Use simulation when neither model fits — multi-skill, non-Poisson, time-varying patience, or routing complexity beyond a single queue.

For most modern contact centers, Erlang-A is the more appropriate model. Erlang-C survives as the simpler approximation and the historical default.

Common failure modes

  • Setting patience equal to the SL threshold. Common shortcut: use T (e.g., 20 seconds) as the patience input. The patience distribution is much longer-tailed than T in nearly every queue. Result: over-stated abandonment and under-staffing.
  • Estimating patience from abandoned-caller waits only. Survival bias: the long-patience callers got served, so their patience never appears in abandon data. The mean patience is longer than the mean abandon-wait suggests.
  • Treating patience as fixed. Patience varies by call type, time of day, queue (a customer holding for retention will wait longer than one holding for routine support), and announcements. A single patience number is a Level 2 simplification.
  • Using Erlang-A on multi-skill queues. Same issue as Erlang-C — single-queue formula breaks for skill-pooled operations.
  • Conflating caller-patience abandonment with technical abandonment. Disconnects, line drops, and routing failures are not patience-driven and shouldn't be in the abandonment number that feeds Erlang-A.

Maturity Model Position

  • Level 1 — Initial (Emerging Operations). Abandonment tracked but not modeled; staffing assumes Erlang-C or rules of thumb. Planning unaware of patience economics.
  • Level 2 — Foundational (Traditional WFM Excellence). Erlang-C drives staffing; abandonment reported as separate post-hoc metric. Erlang-A available in WFM platform but rarely the planning default.
  • Level 3 — Progressive (Breaking the Monolith). Erlang-A used as the planning default for queues with material abandonment. Patience estimated from ACD data per queue. Plans target SL and abandonment jointly.
  • Level 4 — Advanced (The Ecosystem Emerges). Erlang-A supplemented with simulation for complex queues. Patience modeled as call-type and time-of-day specific. Plans report distributional outcomes (SL distribution, abandonment distribution) per the WFM Labs Risk Score™ frame. Patience-weighted value-routing decisions become possible.
  • Level 5 — Pioneering (Enterprise-Wide Intelligence). Patience modeled in real time per customer segment. Erlang-A replaced by learned queue models that adapt to non-stationarity and customer-specific patience.

References

  • Palm, C. (1946). Methods of judging the annoyance caused by congestion. Tele 4, 189–208. The original patience-distribution paper.
  • Garnett, O., Mandelbaum, A., & Reiman, M. (2002). Designing a call center with impatient customers. Manufacturing & Service Operations Management, 4(3), 208–227. The modern Erlang-A (or M/M/c+M) treatment that brought the model into contact-center practice.
  • Mandelbaum, A., & Zeltyn, S. (2007). Service engineering in action: the Palm/Erlang-A queue. Operations Research, 55(2), 165–183.
  • Koole, G. (2013). Call Center Optimization. MG Books. Practical Erlang-A use in contact centers.
  • Aksin, Z., Armony, M., & Mehrotra, V. (2007). The Modern Call Center: A Multi-Disciplinary Perspective. Production and Operations Management, 16(6), 665–688. Empirical patience and abandonment data.
  • Lango, T. (2026). Adaptive: The Workforce Transformation Architecture. Erlang-A in distributional and value-aware planning.

See Also

Interactive tools

  • Erlang Suite — erlangcalculator.wfmlabs.com. Includes Erlang A alongside Erlang C, Power of One, and Day Planner. The right tool when abandonment matters.
  • Power of One — powerofone.wfmlabs.com. Single-agent sensitivity, applicable in both Erlang-C and Erlang-A regimes.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Erlang-A&oldid=745"